Efficient Partitioning of Memory Systems and Its Importance for Memory Consolidation
a. The SNR for two values of for a fixed number of synapses (solid lines: consolidation model, dotted lines: heterogeneous model without interactions). The initial SNR for both models scales as . It then decays as power law () and finally as an exponential for for the heterogeneous model and for for the consolidation model. Three measures of interest are shown in the inset and in the bottom two panels. Inset: crossing time between the SNR of the heterogeneous model and the SNR of the consolidation model as a function of . The heterogeneous model is better than the consolidation model only for very recent memories (stored in the last hours, compared to memory lifetimes of years). b. The SNR scales as in the consolidation model when the SNR decay is approximately a power law (symbols: simulations, line: analytics). The SNR at indicated times is plotted as a function of for three different values of . c. Lifetimes (i.e. time at which SNR = 1) in the consolidation model scale approximately as ( is the fastest learning rate and is the slowest). The memory lifetime is plotted vs for three different values of . synapses evenly divided into stages. Stage has a learning rate .